cera-cranium: a test bed for machine consciousness research€¦ · keywords: cognitive...
TRANSCRIPT
International Workshop on Machine Consciousness 2009
Copyright © 2009, AGI-Network (http://www.agi-network.org). All rights reserved.
CERA-CRANIUM:
A Test Bed for Machine Consciousness Research
Raul Arrabales [email protected]
Agapito Ledezma
Araceli Sanchis Carlos III University of Madrid
Avda. Universidad, 30.
28911 Leganés, Madrid, SPAIN
Abstract
This paper describes a novel framework designed as a test bed for machine consciousness
cognitive models (MCCM). This MCCM experimentation framework is based on a general-
purpose cognitive architecture that can be integrated in different environments and confronted
with different problem domains. The definition of a generic cognitive control system for abstract
agents is the root of the versatility of the presented framework. The proposed control system,
which is inspired in the major cognitive theories of consciousness, provides mechanisms for both
sensory data acquisition and motor action execution. Sensory and motor data is represented in the
proposed architecture using different level workspaces where percepts and actions are generated
thanks to the competition and collaboration of specialized processors. Additionally, this cognitive
architecture provides the means to modulate perception and behavior; in other words, it offers an
interface for a higher control layer to drive the way percepts and actions are generated and how
they interact with each other. This mechanism permits the experimentation with virtually any high
level cognitive model of consciousness. An illustrative application scenario, autonomous explorer
robots, is also reviewed in this work.
Keywords: Cognitive architectures, cognitive modeling, machine consciousness.
1. Introduction
From the point of view of an Artificial Intelligence engineer, most of the existing theories of
consciousness, which typically come from philosophy or psychology, do not provide a fully
plausible explanation of what a conscious being is and how consciousness could be produced in a
machine. Instead, they offer a more or less metaphorical description of consciousness, but not a
model that can be directly implemented in computational terms. Nevertheless, cognitive theories
of consciousness, like Global Workspace Theory (Baars, 1997) or Multiple Draft Model
(Dennett, 1991), have some aspects in common that can be taken as a functional guideline for the
design of at least a partial computational model of consciousness.
Although authors use different names or descriptions, cognitive theories of consciousness
share the assumption that the unity of self produced in conscious beings has its roots in non-
unitary mechanisms. More precisely, it is argued that conscious contents emerge as a result of
ARRABALES ET AL
2
competition and collaboration between specialized processors (Minsky, 1988; Dennett, 1991;
Hofstadter, 1995; Baars, 1997; Shanon, 2008). Different theories offer different explanations or
metaphors regarding the specific way in which these processes of competition and collaboration
take place; however, all theories agree on their highly adaptable and dynamic nature.
Some remarkable examples of machine consciousness implementations inspired in these sorts
of theories are Shanahan’s cognitive architecture (Shanaham, 2005, 2006) and LIDA
(Ramamurthy et al., 2006). Given that the detailed way in which consciousness is produced is not
explained by the aforementioned theories, each existing MCCM take a different approach
regarding the concrete way in which perceptual and action flows are built and managed. There
exists however a common denominator in relation to the underlying mechanism used to perform
low-level cognitive processes: the concurrent collaboration and competition of multiple
specialized processors in a shared workspace. What differs from one implementation to another is
the specific technique applied to orchestrate the collaboration and competition processes.
Additionally, each particular implementation is usually oriented towards specific environments
and problem domains, making it difficult to compare their relative performance. It is our aim to
provide a platform where different high-level cognitive approaches can be tested and compared
with each other. In order to design such a test bed we have developed a generic but configurable
low-level cognitive architecture based on multiple level workspaces. The proposed framework
aims to provide the main functional features of a general-purpose cognitive architecture that can
be used as the base of a higher level computational model of consciousness. Taking into account
the description of conscious content formation described by the main cognitive theories of
consciousness, mechanisms for specialized processors creation, association, combination, and
competition have been implemented as well as appropriate means to regulate these processes.
Adopting a purely cognitive perspective as introduced above does not necessarily imply that
phenomenal aspects of consciousness are neglected in our work. Although our efforts are
specifically focused on functional features of consciousness, phenomenology is expected to be
the main subject of study after all key functional aspects are successfully implemented and tested.
The details of the proposed framework are discussed as follows. Section 2 provides an
overview of CERA-CRANIUM layered design and an overall description of the perception and
action cognitive flows. Section 3 covers the application of the proposed framework to the
particular domain of unknown environment exploration using a mobile robot. Finally,
conclusions and open research issues are discussed in section 4.
2. CERA-CRANIUM Overview
In order to build an efficient framework for the development and testing of cognitive models of
consciousness we have designed and implemented the following main components: CERA, a
control architecture structured in layers, and CRANIUM, a tool for the creation and management
of high amounts of parallel processes in shared workspaces. As we explain below, CERA uses the
services provided by CRANIUM with the aim of generating a highly dynamic and adaptable
perception processes orchestrated by a computational model of consciousness.
2.1 CERA
CERA (Conscious and Emotional Reasoning Architecture) is a layered cognitive architecture
designed to implement a flexible control system for autonomous agents. Current definition of
CERA is structured in four layers (see Figure 1): sensory-motor services layer, physical layer,
3
mission-specific layer, and core layer. As in classical robot subsumption architectures, higher
layers are assigned more abstract meaning; however, the definition of layers in CERA is not
directly associated to specific behaviors:
CERA sensory-motor services layer comprises a set of interfacing and communication services
which implement the required access to both sensor readings and actuator commands. In order to
endow an agent with a full CERA controller system, every sensor must have its corresponding
sensor service; analogously, every actuator must have its corresponding motor service. These
services provide the physical layer with a uniform access interface to agent’s physical (or
simulated) machinery.
CERA physical layer encloses agent’s sensors and actuators low-level representations.
Additionally, according to the nature of acquired sensory data, the physical layer performs data
preparation and preprocessing. Analogous mechanisms are implemented at this level with
actuator commands, making sure for instance that command parameters are within safety limits.
Although sensory information binding does not take place at this level, low-level
contextualization parameters, like relative positions and timestamps, are calculated and annotated
in the physical layer.
CERA mission-specific layer (formerly referred to as instantiation layer) produces and manages
elaborated sensory-motor content related to both agent’s vital behaviors and particular missions
(one mission will typically involve several goals). At this stage is when single contents acquired
and preprocessed by the physical layer are combined into more complex pieces of content, which
have some specific meaning related to agent’s goals. The mission-specific layer can be modified
independently of the other CERA layers according to assigned tasks and agent’s needs for
functional integrity.
CERA core layer, the highest control level in CERA, encloses a set of modules that perform
higher cognitive functions. The definition and interaction between these modules can be adjusted
in order to implement a particular MCCM. In some of our former works (Arrabales et al., 2006,
2007, 2008), we have identified the following core modules: attention, status assessment,
preconscious management, memory management, and self-coordination. Nevertheless, CERA is
designed to allow a custom definition of core modules. The objective of the mentioned modules is
discussed below as well as the mechanisms they use to modulate the way CERA lower layers
work.
.
Agent’s Body
Physical Layer
Mission-specific Layer
Core Layer
Sensors
Actuators
CERA World
Sensory-accessible
environment
Reachable
environment
Sensor
Services
Motor
Services
Figure 1. CERA cognitive architecture layered design.
ARRABALES ET AL
4
In order to perform an experiment using the proposed test bed, four variables have to be assigned
a particular value: cognitive model of consciousness to be tested, physical (or simulated) agent,
assigned mission, and environment. This so-called instantiation process involves the definition of
CERA core layer modules, implementation of interfaces for the particular agent’s sensors and
actuators (CERA sensory-motor services layer), and definition of mission-specific routines
(CERA mission-specific layer). For the particular case of MCCM comparative study,
environment, mission, and agent have to remain constant; therefore, the only changes required in
order to test different cognitive models of consciousness have to be made within the CERA core
layer.
Physical and mission-specific layers are characterized by the inspiration on cognitive theories
of consciousness, where large sets of parallel processes compete and collaborate in a shared
workspace in the search of a global solution. Actually, a CERA controlled agent is endowed with
two hierarchically arranged workspaces which operate in coordination with the aim to find two
global and interconnected solutions: one is related to perception and the other is related to action.
In short, CERA has to provide an answer for the following questions continuously:
1. What must be the next content of agent’s conscious perception?
2. What must be the next action to execute?
Typical robot control architectures are focused on the second question while neglecting the first
one. Here we argue that a proper mechanism to answer the first question is required in order to
successfully answer the second question in a human-like fashion. Anyhow, both questions have to
be answered taking into account safety operation criteria and the mission assigned to the agent.
Consequently, CERA is expected to find optimal answers that will eventually lead to human-like
behavior. As explained below, CRANIUM is used for the implementation of the workspaces that
fulfill the needs established by the CERA architecture.
2.2 CRANIUM
CRANIUM (Cognitive Robotics Architecture Neurologically Inspired Underlying Manager)
provides a software library in which CERA can execute thousands of asynchronous but
coordinated concurrent processes. In addition to the design guideline based on the main cognitive
theories of consciousness, CRANIUM is also inspired by the way brain works from the systems-
level point of view, where specialized regions process information coming both from the senses
or from other specialized regions. According to the global access hypothesis (Baars, 2002), neural
connections between specialized areas make possible the emerging global coordination.
A CRANIUM workspace can be seen as a particular implementation of a pandemonium, as
described in (Dennett, 1991), where demons compete with each other for activation. Each of these
demons or specialized processors is designed to perform a specific function on certain types of
data. At any given time the level of activation of a particular processor is calculated based on a
heuristic estimation of how much it can contribute to the global solution currently sought in the
workspace. The concrete parameters used for this estimation are established by the CERA core
layer as explained below. As a general rule, CRANIUM workspace operation is constantly
modulated by commands sent from the CERA core layer.
In the proposed framework we use two separated but connected CRANIUM workspaces
integrated within the CERA architecture. The lower level workspace is located in the CERA
physical layer, where specialized processors are fed with data coming from CERA sensor
5
services. The second workspace, located in the CERA mission-specific layer, is populated with
higher-level specialized processors that take as input either the information coming from the
physical layer or information produced in the workspace itself (see Figure 2). The perceptual
information flow is organized in packages called single percepts, complex percepts, and mission
percepts. The details about these constructs are explained in the next subsections.
In addition to the bottom-up flow involving perception processes, a top-down flow takes place
simultaneously in the same workspaces in order to generate agent’s actions. Physical layer and
mission-specific layer workspaces include single actions, simple behaviors, and mission
behaviors (see Figure 3). The detailed way in which these motor representations are managed is
also covered in the following subsections.
One of the key differences between CERA-CRANIUM bottom-up and top-down flows is that
while percepts are being iteratively composed in order to obtain more complex and meaningful
representations, high level behaviors are iteratively decomposed until a sequence of atomic
actions is obtained. As described below, there are different types of specialized processors that
can be implemented and associated with CRANIUM workspaces. At first glance, it seems that
specialized processors either take percepts or behaviors as input. Nevertheless, some specialized
processors are designed to generate behaviors as a function of received percepts. For instance,
high-priority reactive responses are rapidly generated in the physical layer, typically without any
intervention from the upper layer. If a complex percept indicating a physical threat appears in the
physical workspace, the specialized processor in charge of detecting this sort of threat will be
activated and it will generate a reactive simple behavior as a response (see Figure 4). The
obtained evasive simple behavior will be selected and the corresponding sequence of actions will
be executed by the CERA sensory-motor services layer.
CERA
Core
Layer
CERA M-S Layer CERA Physical Layer
Sensor
Service
Sensor Service
Single
Percepts
…
CRANIUM
Workspace
Complex
Percepts
…
CRANIUM
Workspace
Mission
Percepts
Sensor
Preprocessors
…
Specialized
Processors
Sensor
Service
CERA S-M Sensors
…
Percept
Aggregators
CERA
Core
Layer
CERA M-S Layer CERA Physical Layer
Motor
Service
Motor
Service
Single
Actions
…
CRANIUM
Workspace
Simple
Behaviors
…
CRANIUM
Workspace
Mission
Behaviors
Action Preprocessors
… Specialized Processors
Motor
Service
CERA S-M Actuators
…
Action
Planners
Figure 2. CERA-CRANIUM bottom-up flow: perception.
Figure 3. CERA-CRANIUM top-down flow: behavior generation.
ARRABALES ET AL
6
A CRANIUM workspace provides a shared and global access working memory where data can
be read or written by any associated specialized processor. In the particular case depicted in
Figure 4, which has been simplified for the sake of clarity, three single percepts are generated and
added to the physical workspace as a consequence of an impact detected almost simultaneously in
three contiguous bump panels of a mobile robot (this particular example is studied in detail in the
next section). The percept aggregator processor reads these single percepts and builds a new
complex percept out of their data. The concrete set of single percepts selected in order to form the
new complex percept is determined thanks to the application of multiple context criteria. Once
the new complex percept is generated and published in the workspace, a reactive processor can
read it and detect a condition in which a quick reactive response is required. If such a condition is
met, the reactive processor is able to build a simple behavior designed to diminish or prevent the
consequences of the detected undesired situation. The presence of the reactive simple behavior in
the workspace triggers the activation of the action planner specialized processor, which in turn
will produce the corresponding sequence of single actions.
As discussed in the former example, there are various types of specialized processors. All
these processors work as asynchronous independent programs that are able to subscribe to a
workspace, read certain data types from it, perform a particular processing, and then submit new
elaborated data back to the workspace, where it becomes available for the rest of specialized
processors. Basically, a CRANIUM workspace interacting with its associate processors
constitutes a blackboard system (Nii, 1986), where CERA plays the role of the blackboard control
shell. Most important CRANIUM processor types are briefly described as follows (an exhaustive
list of all processor types is out of the scope of the present work):
Sensor preprocessors collect raw sensory data that appears in the physical workspace and build
single percepts by combining sensor readings with contextual information that can be associated
to them. The generated single percepts are automatically sent to the same workspace. In order to
perform this task, sensor preprocessors also retrieve system information available in the
workspace, like limbs relative positions and timing.
Action preprocessors prepare atomic actions generated by action planners (another type of
processor) to enter the execution cycle. Basically, action preprocessors build the so-called single
action constructs which include contextual data about actions. For instance, every single action
encloses the exact timestamp corresponding to the moment it was created (planned) and also the
actual timestamp assigned for execution by the CERA action dispatcher. This information is used
to abort the execution of actions that have been queued for too long. Proprioceptive sensory data
is also included in order to adapt actions to the current position of the actuators.
Contact
Sensor Service
Drive Service
Single
Percepts
Simple
Behavior
Motor
Controllers
Crash!
Contact
Sensors
WORLD AGENT CERA S-M
CERA Physical Layer
Actions
Complex
Percept Reactive Processor
Workspace
Action
Planner
Percept
Aggregator
Figure 4. Simplified scheme of the reflex mechanism in a mobile robot controlled by CERA-CRANIUM.
7
Percept aggregators are the processors in charge of building complex percepts out of
interrelated single percepts. While single percepts represent atomic sensory information, complex
percepts are more elaborated and meaningful combinations of the former. Multiple context
criteria can be considered in the formation of complex percepts. Consequently, different types of
percept aggregators will focus on different parameters for the selection of the single percepts they
combine. As soon as complex percepts are generated, percept aggregators send them to the
workspace, where other processors could use them for further aggregation processes.
Reactive processors are typically located in the physical layer in order to provide a quick
response to stimuli that are considered harmful or highly undesired for the agent. These
processors monitor the generated single and complex percepts looking for particular unsafe
conditions. If such conditions are met, the processors build simple behaviors intended to mitigate
the detected risk.
Action planners are processors able to take a behavior as input and generate the corresponding
sequence of atomic actions that will lead to behavior completion. Thanks to action planners all
active behaviors in a workspace are processed and the corresponding action sequences are
submitted to be eventually executed.
Sensory Predictors monitor a particular source of sensory information incessantly, interpreting it
as a continuous signal that can be predicted in the short term. When the sensory input under
analysis differs significantly from the prediction, these processors build a mismatch complex
percept that is placed in the corresponding workspace. Mismatch complex percepts are used to
concentrate attention on unusual perceptions.
Having a shared workspace, where sensory and motor flows converge, facilitates the
implementation of the multiple feedback loops required for adapted and effective behavior. The
winning simple behavior is continuously confronted to new options generated in the physical
layer, thus providing a mechanism for interrupting behaviors in progress as soon as they are no
longer considered the best option. In general terms, the activation or inhibition of perception and
behavior generation processes is modulated by CERA according to the implemented cognitive
model of consciousness. Figure 5 shows a schematic representation of typical feedback loops
produced in the CERA architecture. These loops are closed when the consequences of actions are
perceived by the agent, triggering adaptive responses at different levels.
Curve (a) in Figure 5 represents the feedback loop produced when an instinctive reflex is
triggered as in the example depicted in Figure 4. Figure 5 curve (b) corresponds to a situation in
which a mission-specific behavior is being performed unconsciously. Finally, curve (c)
CERA M-S Layer CERA Core Layer CERA Physical Layer CERA S-M WORLD
(a)
(b)
(c)
Figure 5. Different feedback loops produced in the CERA-CRANIUM.
ARRABALES ET AL
8
symbolizes the higher level control loop, in which a task is being performed consciously. These
three types of control loops are not mutually exclusive; in fact, same percepts will typically
contribute to simultaneous loops taking place at different levels.
As explained above, the implementation of CERA-CRANIUM computational model has
strong requirements in terms of concurrency and asynchronous input/output. The software
architecture designed to address this issue is described in the next subsection.
2.3 Software Architecture
The concepts about CERA-CRANIUM described above refer to the cognitive-level architecture.
However, building such a system also requires an underlying well designed software architecture.
A good software engineering strategy will allow us to have a robust and extensible artifact that
can be easily modified, enhanced and reused. Additionally, performance and scalability are
factors that cannot be ignored due to the large computational demands correlated with a high
number of specialized processors to be executed concurrently. Bearing these considerations in
mind, as well as the requirement of a powerful physics simulator, the software development
platform selected for the implementation of CERA-CRANIUM is Robotics Developer Studio
2008 (Microsoft, Corp., 2008).
The Robotics Developer Studio runtime is based on two key components: the CCR or
Concurrency and Coordination Runtime (Richter, 2006), which we use for asynchronous
programming and specialized processors concurrency management; and the DSS or Decentralized
Software Services (Nielsen and Chrysanthakopoulos, 2006), which provide us with a framework
for implementing a light-weight distributed service-oriented architecture. Applying the service
orientation paradigm (Singh and Huhns, 2005), each CERA layer has been defined as an
independent service that, if needed, can be executed in a separate machine. Consequently, the
communication between layers is implemented using the DSS protocol (DSSP). Each CRANIUM
workspace allocates at least one managed high-performance thread pool that dispatches
specialized processors tasks across all available CPUs. CRANIUM thread dispatching mechanism
and asynchronous I/O coordination patterns are adjusted by CERA core layer commands. Figure
6 depicts a simplified view of main software architecture components and their communication
scheme (circles between modules indicate asynchronous communication implemented using CCR
task ports).
Agent sensory-motor
machinery
.Net Framework
CCR
CRANIUM
DSS
CERA Physical
CERA Sensory-
Motor Services
.Net Framework
CCR
CRANIUM DSS
CERA Mission-specific
.Net Framework
CCR
DSS
CERA Core
MCCM Configuration
DSSP DSSP
Figure 6. CERA-CRANIUM software architecture.
9
2.4 Knowledge Representation
Knowledge representation is one of the key problems in artificial general intelligence, and so is in
the particular field of machine consciousness. Any machine consciousness model should provide
a satisfactory account for world knowledge representation and symbol grounding (Haikonen,
2007). In CERA-CRANIUM, sensory and motor data is iteratively processed inside the
workspaces and across layers in order to build higher level meaningful knowledge about the
world. Raw sensory data coming directly from the sensors is initially processed by specific sensor
preprocessors in the physical layer workspace. These preprocessors build single sensor data
representations called single percepts. Single percepts are integrations of mono-modal sensor data
packages and their associated contextualization parameters. Basically, contextualization
parameters characterize the perceived stimulus in terms of relative position and time of the
sensing event (see Figure 7). CERA physical layer encloses a number of modules designed to
keep track of physical variables. For instance, the timer module implements a precision clock that
represents the age of the agent down to a resolution of 1 millisecond. In addition, the
proprioception module calculates the position of exteroceptive sensors. These parameters are used
by the sensor preprocessors to calculate the relative location of the percept being acquired by the
corresponding exteroceptive sensor. As shown elsewhere (Arrabales et al., 2009a), additional
contextualization parameters can be established in CERA in order to obtain more accurate and
selective perceptions.
Following Aleksander’s notation for axioms of neuroconsciousness (Aleksander and Dunmall,
2003), where S is the sensory accessible world and δSj represents a minimal percept. A necessary
requirement for perception is that such a minimal percept must have a correlated agent’s internal
physical state: N(δSj). According to this notation, j represents the relative location, that is, the
encoding of the location where the percept has been originated from the point of view of the
observer organism. Consequently, N(S) is the entire internal representation of the world built by
the agent. In CERA physical layer, single percepts produced by sensor preprocessors include
relative contextual information that we call J (J also include representations to encode relative
locations called j referent vectors). In other words, single percepts can be referred to as N(δSJ),
where J is a set of contextual parameters including j (relative position) and t (timestamp).
Therefore, N(S) is the union of the N(δSJ) representations produced by CERA (see Equation 1). In
CERA, the representation of relative contextual parameters, J, is not encoded using neural
networks, but explicit representations in the form of geometrical vectors or integer variables (as
explained in the example discussed in next section).
Sensor
Single Percepts
Sensor
Preprocessors
Sensor Readings
Timer
Proprioception
Sensor
Service
Agent
CERA sensory-
motor services CERA physical
layer
Percept
Aggregators
S (World)
events δSj N(δSj)
N(δSJ) j t
Complex Percept
M(SCJ)
(1)
Figure 7. Single percepts generation in CERA physical layer.
ARRABALES ET AL
10
All single percepts are produced within the physical layer CRANIUM workspace, where they
become accessible immediately to all physical layer specialized processors. Percept aggregator
processors are able to combine two or more single percepts available in the workspace and build a
so-called complex percept. Complex percepts can be mono-modal or multi-modal representations
depending on the modality of the single percepts they come from. Basically, they constitute more
elaborated representations of the world that have been assembled as a result of the application of
certain contextualization rules.
Similarly to single percepts, newly generated complex percepts are immediately published in
the CERA physical workspace, and they are also sent to the mission-specific layer workspace.
This means they become available to the specialized processors of both layers. Although current
implementation does not include processors that combine several complex percepts into one
larger complex percept, this feature could be added just by defining the corresponding specialized
processors. Nevertheless, a composed J-index (CJ) is always calculated for each new complex
percept. Percept aggregators combine the J-indexes coming from the original single percepts and
build an integrated CJ-index that represents the contextualization parameters of the whole
complex percept being formed. As single percepts are J-indexed, they can be grouped in terms of
context criteria; for instance, a particular specialized processor might select all single percepts
sensed 10 seconds ago in the left hand side of the agent and build the corresponding complex
percept. According to this definition, complex percepts can be referred to as M(SCJ), a subset of
agent’s internal world representation (see Equation 2).
Problems can be encountered when a percept aggregator is building a new complex percept out of
contradictory single percepts. This situation can be caused by sensor noise or malfunctioning
hardware. Single percepts corresponding to different sensors but associated by a common context
could provide contradictory data. In that case, some strategies can be applied in order to build a
meaningful complex percept that successfully integrates all the data from the original single
percepts. One option is to assign levels of confidence both to the sensory data and contextual
parameters obtained by sensors. Other complementary option is to generate mismatch complex
percepts that will raise core layer attention towards the unexpected situation.
CERA mission-specific layer hosts a workspace in which complex percepts received from the
lower layer become the input of mission-specific processors. Once more, mission-specific layer
percepts are sent both to the same layer workspace and to the upper layer (CERA core layer in
this case). Mission percepts could have been generated directly using one single CRANIUM
workspace where single and complex percepts could be also included. However, using separated
workspaces allow us to decouple physical agent specific processors and mission-specific
processors. Some examples of mission-specific processors are briefly discussed in the next
section.
Motor data representation in CERA-CRANIUM is analogous to the sensory data
representation explained above. Atomic actions are defined as δBI (Arrabales et al., 2007), being I
the referent that indicates the parameters of movement (like direction and speed). M(BCI)
represents a generic behavior, which is defined as a sequence of atomic actions. The CI notation
refers to the integrated CI-index, which is the final context expected to be reached when the
behavior is completed. For instance, if M(BCI) refers to a U-turn movement, CI-index will indicate
the final position of the agent once the U-turn is completed. If the U-turn behavior is decomposed
into atomic actions, a sequence of I-indexes corresponding to the sequence of steps required to
(2)
11
perform the U-turn will represent the intermediate positions and speeds. Applying the same
notation as used for percepts, N(δBI) refers to the representation hold in CERA physical layer for
agent’s atomic actions. Analogously to the bottom-up processing carried out by sensor
preprocessors, action preprocessors perform a top-down processing in order to build single
actions out of atomic actions by including contextual information about time and relative position
(see Figure 8).
CERA action dispatcher manages an execution queue in which active single action sequences
wait for execution. Single actions constructs include some parameters that can be used by the
dispatcher to decide how to manage the queue. For instance, actions derived from highest priority
simple behaviors are executed in the first place.
Mission-specific behaviors are generated in the corresponding CERA layer, and then
submitted to the physical layer where they are decomposed into a sequence of simple behaviors.
The activation of mission-specific behaviors is driven by the application of mission goals and
commands sent from the core layer. In general, the operation of CERA physical and mission-
specific layers is modulated by workspace commands sent from the core layer.
2.5 Workspace Modulation
CRANIUM workspaces are not passive short-term memory mechanisms. Instead, their operation
is affected by a number of workspace parameters that influence the way the pandemonium works.
These parameters are set by commands sent to physical and mission-specific layers from the
CERA core layer. In other words, while CRANIUM provides the neural-like mechanism for
specialized functions to be combined and thus generate meaningful representations, CERA
establishes a hierarchical structure and modulates the competition and collaboration processes
according to the model of consciousness specified in the core layer. This mechanism closes the
feedback loop between the core layer and the rest of the architecture: core layer input (perception)
is shaped by its own output (workspace modulation), which in turn determines what is perceived.
All theories of consciousness differentiate between implicit and explicit processing (Atkinson
et al., 2000). In CERA-CRANIUM all sensory-motor contents being processed in the workspaces
are by default implicit or unconscious contents. The selection of a reduced subset of contents
which will become available for explicit reasoning is carried out by the competition of both
specialized processors and percepts. CERA-CRANIUM provides a mechanism to modulate these
competition processes by means of commands sent from the core layer, where the cognitive
model of consciousness to be tested is implemented. Specialized processors, behaviors (simple
behaviors and mission-specific behaviors) and percepts (single percepts, complex percepts, and
mission specific percepts) are assigned an activation level by the workspace. Activations levels
are highly dynamic variables that are being constantly updated by the workspace. The role of
these activation levels is twofold: on the one hand, percepts with a very low activation level are
Actuator
Atomic Actions
Action
Preprocessors
Single Actions
Timer
Proprioception
Motor
Service
Agent
CERA sensory-
motor services
CERA physical
layer
Action
Planners
S (World)
Action δBi
N(δBi) N(δBI) j t
Simple Behaviors
Action
Dispatcher M(BCI)
Figure 8. Actions generation in CERA physical layer.
ARRABALES ET AL
12
not processed any further, thus saving the limited computational resources; on the other hand,
percepts with the highest activation levels are iteratively processed until a winner is selected as
conscious content. The operation of specialized processors is also affected by their activation
levels. Processors with lower activation levels are less likely to be executed as they are assigned
less priority in the workspace thread dispatching.
Activation levels for specialized processors and percepts are calculated as a function of
multiple parameters, being contextualization criteria the most important ones. Context commands
are messages sent to the workspaces specifying a J-index. This contextual J-index establishes the
context criteria (like time and relative location as discussed above) that have to be activated in the
workspaces. For instance, a J-index might refer to the particular segment of agent’s visual field of
view that fall between 8º and 14º. A context command sent with such a J-index would cause the
increase of the activation level of those percepts with J-indexes indicating they have been sensed
close to that particular position. The assigned activation level is inversely proportional to the
distance between context command J-index and percept CJ-index. When a CRANIUM
workspace receives a context command, activation levels of all percepts and processors inside the
workspace are automatically recalculated. The distance between the specified contextual J-index
and existing percepts J-indexes is calculated, and percept activations are assigned accordingly.
The level of activation of processors is assigned in terms of the input they can process (activation
of their potential input).
As discussed in the example presented in the next section, active contexts are typically
established in the core layer taking into account agent’s goals and feedback obtained from lower
layers. Percept activation is also based on the match/mismatch/novelty mechanism proposed by
Haikonen (2007). For instance, a mismatch percept will be initially assigned a high activation
level because it might represent part of an unexpected situation that may require conscious
attention. Once the mismatch signal reach the core layer, the MCCM can induce a contextual bias
in the lower CERA levels by sending a context command specifying a J-index directed towards
the unexpected percept.
The contextual bias induced by the core layer determines the percepts that are formed.
Consequently, the perception process is a highly active mechanism rather than a passive data
retrieving system. As behaviors are also assigned an activation level, behavior generation is also
affected to a great extent by active contexts. At any given time, a number of possible behaviors
are generated in the workspaces; however, only those with the highest activation levels are likely
to be selected and finally executed. The calculation of activation levels for behaviors is also based
on the distance between the contextual J-Index and behaviors CI-Indexes. In other words, those
behaviors which are directed to the same location as the active contexts will be selected. In fact,
the contextual J-index is not only used for selecting existing behaviors, but also to generate new
behaviors directed to current focus of attention. The application of these mechanisms is illustrated
in the next section with a domain-specific example. The implementation of a very basic core layer
is also discussed.
3. Application in Autonomous Robot Exploration
As discussed above, our proposed framework has been designed to be used as a test bed in
multiple scenarios. Typical application environments for machine consciousness research include
autonomous robots and virtual agents; see for instance (Holland, 2007) and (Goertzel, 2008). In
13
the following we briefly introduce an instantiation of the CERA-CRANIUM framework in the
domain of unknown environment exploration and mapping using autonomous mobile robots.
3.1 Experimental Setting
Preliminary experiments have been developed using real and simulated Pioneer 3DX robots
equipped with front and rear bumper arrays and a ring of eight forward ultrasonic transducer
sensors (see Figure 9).
As the mapping task is simplified to a two-dimensional grid, J-indexes are composed of j-
referent vectors that represent relative positions using just two coordinates, being (x,y) = (0,0) the
subjective point of view of the robot. In this particular case of unknown environment exploration,
spatial contexts have been defined in order to estimate optimal headings during robot navigation.
A specific version of CERA mission-specific layer has been coded with the aim of representing
the particular complex percepts that are required for the mapping mission. Concretely, sonar
single percepts that represent obstacles are combined firstly into complex percepts, and secondly
into mission-specific percepts that represent walls and corridors.
Figure 9. Simulated and real Pioneer 3DX robots.
Based on (mission-specific) internal map representations (see Figure 10), which are continuously
updated with new sensed percepts, CERA core layer calculates an adaptive workspace
modulation response as indicated by the implemented MCCM.
Figure 10. Mission-specific percept representing the map.
Each sensor provides different measurements of j-referents. As explained above, sonar range data
is used for building mission percepts that represent walls or corridors. Given that sonar sensors
usually provide noisy readings in real environments, contact sensors are also used for robustness
ARRABALES ET AL
14
(especially when maneuvering too close to obstacles). For the sake of conciseness, in the
following we only explain how single and complex percepts are built out of robot bumpers
sensory information. For a detailed description of sonar percepts formation see (Arrabales et al.,
2009a). Pioneer 3DX bumper arrays consist of five points of contact sensing arranged at angles
around the robot (see Figure 11).
Figure 11. Pioneer 3DX frontal bumper array.
The j-referents of single percepts generated when the bump panels are pressed are calculated
using equation 3. Where BR is the bump panel radius (or the distance from the origin of the robot
spatial reference system to the bumper contact surface), and BA is the bump panel angle to the
front of the robot (bump panels are located at angles -52º, -19º, 0º, 19º, and 52º).
Additionally, two more vectors are calculated to be associated to a bumper single percept: the
left-j referent and the right-j referent vectors (see Figure 12). These two vectors represent the
dimensions of the percept (the width assigned to the collision). All these referent vectors plus the
contact timestamp are used to build the J-index of the bumper single percept. The combination of
the J-Index and the sensory data provided by the sensor constitutes the N(δSJ). In the case of
bumper single percepts sensory data is just the indication of a bump panel press or release. Other
sensor modalities will include other types of data like range measurements or image bitmaps.
Figure 12. Referent vectors calculated to build the J-index of a bumper single percept.
(0,0)
-52º
-19º
0º
19º
52º
BAb5
BR
(3)
b1
b2
b3 b4
b5
Impact
j referent Left-j referent
Right-j referent
N(δSJ)
15
Using time and relative location parameters as contextualization criteria single percepts can
be related and associated forming complex percepts. For example, if the bumper service from the
CERA sensory-motor services layer reports contacts in bump panels b2, b3, and b4
simultaneously (see Figure 13), three single percepts are created by bumper sensor preprocessors
in CERA physical layer. Then, these three single percepts can be associated by temporal and
location active contexts by a percept aggregator.
Figure 13. Calculation of the J-index of a bumper complex percept.
The newly created complex percept is assigned a new CJ-index, which is obtained as a
combination of single percepts J-indexes (note that Figure 13 solid lines depict the complex
percept CJ-index, while dashed lines represent the referent vectors of the old single percepts). As
b2, b3, and b4 are located side by side, the CJ-Index of the new bumper complex percept will
have as left-j referent the left-j referent of single percept triggered by b2; analogously, complex
percept right-j referent will be the same as right-j referent from single percept triggered by b4.
The way in which the CJ-Index of a complex percept is calculated depends on the nature
(shape, dimensions, etc.) of the single percepts that take part in the context that gave place to it,
and has to be calculated by the corresponding percept aggregator. The composition of CJ-indexes
is trivial when all single percepts belong to the same modality. However, the composition can be
much more complicated when different modalities are involved. See (Arrabales el al., 2009a) for
details about the calculation of CJ-indexes for multimodal complex percepts.
Motor capabilities of the Pioneer 3DX robot are based on a two-wheel differential drive
system. Robot movement control has been greatly simplified in the current setting, considering
only two atomic actions: rotate in place and move straight. This means that any high level
behavior will ultimately be represented in terms of sequences of these single actions. Attending to
the relative direction specified by the active contextual J-index, an angle parameter is calculated
for the rotate in place operation in order to set the robot heading towards the location that “called
the robot’s attention”. Also a speed parameter is calculated as a function of the distance to the
object.
In addition to the global exploration behavior, which is driven by the top-down contextual J-
index, local obstacle avoidance behaviors are also implemented thanks to some mission-specific
processors like Nearest-Obstacle-Detector and Possible-Impact-Detector. The first one monitors
sonar percepts and creates mission percepts indicating the position of nearest obstacle; the second
one reads sonar complex percepts and current active behavior in order to calculate the possibility
of a collision. The output of these processors is in turn monitored by a reactive processor, which
Impact
Left-j referent
Right-j referent
j referent
M(SCJ)
ARRABALES ET AL
16
will generate a high priority simple behavior for obstacle avoidance only when there is an
obstacle close to the robot and the robot is approaching it (instead of steering away from it).
3.2 CERA Core Layer Design
Taking into account that minimizing exploration time is a requirement of the autonomous
mapping mission, attention should be focused on those areas which have not been previously
visited. This functionality could be implemented as some sort of attentional mechanism located in
the CERA mission-specific layer. However, one of the key features of CERA-CRANIUM is that
higher cognitive abilities, like attention, are not directly implemented as mission-specific
procedures. In fact, higher cognitive abilities should be problem domain independent, and
therefore implemented in the core layer. This means that the operation of the core layer is
directed by meta-goals instead of mission-specific goals. Here is where the specific definition of
a MCCM comes into play, because meta-goals are defined in terms of the particular cognitive
model being implemented. For instance, discovering abstractions (detecting an invariant in a
variance) is a possible definition for a mission-independent meta-goal.
The definition of goal types at different CERA layers can help illustrate the role of the core
layer and the implemented cognitive model. Feedback loops depicted in Figure 5 can also be
expressed in terms of different level goals: (a)-type loops or reflexes are the expression of
physical level goals (basic-goals). These goals are assigned the higher priority and can trigger
behaviors without the intervention of higher layers. (b)-type loops trigger behaviors directed to
provide full or partial solutions to the specific problem domain (mission-goals). Finally, (c)-type
loops are generated as a mean to achieve meta-goals. The precise definition of meta-goals and
how they affect the operation of CRANIUM workspaces is defined by the cognitive model
implemented in the core layer. While the available set of mission-goals defines the possible
functionality of the robot, meta-goals shape the overall resulting behavior. In other words, CERA
core layer provides the mean to orchestrate the operation of the whole control system using
models for higher level cognitive features like emotions, attention, imagination, etc. Examples of
the implementation of these features using CERA-CRANIUM are briefly described below.
According to the implemented MCCM, the CERA core layer is intended to periodically
calculate a contextual J-index that represents current cognitive region of interest for the robot. In
fact, this J-index represents the bias (workspace modulation) that core layer will induce in lower
CERA layers. The input data processed by the cognitive model in order to calculate an adaptive
workspace modulation consists of the set of mission and complex percepts received from the
mission-specific layer (see Figure 14). These percepts provide information about contexts (CJ-
indexes), activation, match/mismatch/novelty signals, mission-goals level of accomplishment,
behaviors being currently executed, etc.
Physical Layer
Mission Percepts, Complex percepts,
Mismatch Percepts,
Novelty Percepts,
…
M(SCJ)
…
CRANIUM
Workspace
…
CRANIUM
Workspace
M-S Layer Core Layer
Workspace
Commands
Current
Model
State
Contextual
J-Index
Calculation
Cognitive Model
Meta-goals
Rule-based
system
Figure 14. Generic implementation of a cognitive model in the CERA core layer.
17
The main task of the cognitive model implementation is to iteratively calculate an adaptive
contextual J-Index. In the unknown environment exploration example, J is bi-dimensional,
because only time and relative location are considered. Nevertheless, contexts could be defined
using additional properties; consequently, J-Indexes can be defined as n-dimensional constructs.
From the point of view of the Global Workspace Theory, the core layer generates the
contexts that outline the border of the metaphorical spotlight. From the point of view of the
Multiple Draft Model, the core layer is in charge of selecting the winning version (reduced set of
percepts) which will become the explicit content of the mind. In short, while physical and
mission-specific layers provide the required environment for the composition of implicit percepts,
the execution of the model implemented in the core layer induces a modulation which directs the
generation and selection of explicit percepts. CERA core layer is designed to host any model able
to generate a J-index as a function of the incoming percepts. Figure 14 represents a simplified
implementation of a generic MCCM. In the illustration, a rule-based system takes incoming
percepts and current state as input and generates an adaptive contextual J-index. Rules are
expected to be defined based on the model meta-goals, and state can be maintained as indicated
by rules. Meta-goals are defined as mission-independent goals, that is, they are defined
exclusively in terms of MCCM parameters. For instance, if the model considers emotions, a
possible meta-goal can be to keep a positive emotional state.
In the case of our autonomous explorer robot, a simple model has been implemented
considering emotions. A reduced set of basic emotions has been considered, as well as their
associated rules that contribute to the final calculation of an adaptive contextual J-index. For
instance, curiosity is defined as an emotion that directs attention toward selected contents.
Therefore, an incoming CJ-index associated with a novelty percept will trigger the curiosity rule,
and contribute to direct the next contextual J-Index towards the novelty CJ-index. Novelty
percepts are generated in the mission-specific layer by specialized processors able to perform
scans over map mission percepts. The example illustrated in Figure 10 shows a map mission
percept and also a 22.5º inclination j referent vector generated as a novelty percept. The
application of the curiosity rule will likely produce a contextual J-Index in the core layer that will
make the robot move in that particular direction. As the robot moves new single percepts are
generated, then combined into complex percepts, map mission percepts, match/mismatch/novelty
percepts, etc. Using just the selected percepts produced in the mission-specific workspace
(explicit content), the MCCM implemented in the core layer will issue a new context command,
thus closing the explicit control loop.
3.3 From Sensory Data to Qualia
How the proposed CERA-CRANIUM framework can account for qualia? Adopting an
engineering approach, as proposed by Haikonen (2008), we believe to inspect the world through
explicit percepts that appear in the system as qualia. Consequently, conscious perception is not
possible without qualia. In terms of CERA, lower layers provide the required mechanisms for
sensory data acquisition, processing, composition, and selection, which take place covertly. Only
a winning selection of complex or mission percepts is overtly available for explicit reasoning.
These explicit percepts, although grounded and adapted to the reality thanks to the CERA
underlying mechanisms, do not represent the real qualities of the outside world, but an impression
created by the multi-layer sensory system. In short, the input of the core layer is not built upon
external stimuli, but based on lower layers reactions to these stimuli. At this level, the functional
role of explicit percepts and their associated CJ-indexes is to provide the required information to
ARRABALES ET AL
18
create the illusion that they directly represent the outside world qualities. However, as pointed out
by Haikonen (2008), the qualities of generated qualia do not necessarily match with physical
world properties.
CERA architecture is designed to integrate exteroceptive and proprioceptive sensing in such a
way that J-indexes can be estimated for each percept as illustrated above. Thanks to the spatial
localization properties derived from the j referent vectors, percepts can be processed as if they
were located in the outside world. We argue that this ability to process percepts as if they were
qualities of the outside world, instead of inspecting directly the sensory data, can be the base for
the creation of qualia in the machine. For instance, the CERA implementation for the robot
exploration task is able to build sonar single percepts out of sonar transducers range data. These
percepts are representations of qualities of the outside world based on sonar sensor current
operation parameters, robot position, and the actual range data (see Figure 15).
Figure 15. Single sonar percept representation.
4. Conclusions and Future Work
In this work we have presented a framework that enables the comparative study of different
cognitive models of consciousness. As illustrated in the robot exploration application scenario
that has been analyzed, particular implementations of CERA layers can be combined in an
instantiation of the proposed framework. Having different implementations of CERA layers
would enable the development of experiments like the following: evaluation of the same MCCM
in different environments, evaluation of the same MCCM in the same environment but using
different agents and different missions, and comparative study of different MCCMs by
confronting them to the same environment and mission (this would only require modifications in
CERA core layer).
The ultimate aim of these sorts of experiments is to discover what workspace modulation
techniques and what cognitive models are best suited to produce conscious-like behaviors and
rich adaptive perception. Furthermore, the proposed experimentation framework can be extended
with classical artificial intelligence search and optimization approaches in order to either improve
existing models or even generate completely new cognitive models of consciousness.
In addition to the former considerations, another benefit to take into account is the existing
decoupling between the control architecture itself and the physical or simulated agent being
controlled. Moving from one particular agent to another would only imply to adapt CERA
sensory-motor services and physical layers to new sensors and actuators, while keeping the rest of
the architecture unaffected by the change. Obviously, if new sensor modalities are added that
could be specifically used in the mission-specific cognitive processing, CERA mission-specific
Sonar transducer
Ultrasonic beam
(three-dimensional cone)
Left-j referent
Right-j referent
j referent vector
(| j | = range measurement)
19
layer would need to be enhanced with this additional capability. The same implication applies in
analogous terms for new actuators and their associated agent physical abilities.
Enhancing CERA-CRANIUM with flexible mechanisms for long-term memory and learning
are the challenges we are currently facing. In order to have a noise-free and rich experimentation
environment for the development of these cognitive capabilities we are currently developing an
instantiation of CERA adapted for the control of video game characters (Arrabales et al. 2009b).
Acknowledgements
This research has been supported by the Spanish Ministry of Science and Innovation under
CICYT grant TRA2007-67374-C02-02.
References
Aleksander, I. and Dunmall, B. 2003, "Axioms and Tests for the Presence of Minimal
Consciousness in Agents", Journal of Consciousness Studies, vol. 10, no. 4-5.
Arrabales, R., Ledezma, A., and Sanchis, A. 2009a, "A Cognitive Approach to Multimodal
Attention", Journal of Physical Agents, vol. 3, no. 1, pp. 53-64.
Arrabales, R., Ledezma, A., and Sanchis, A. 2009b, "Designing Human-like Video Game
Synthetic Characters through Machine Consciousness". In press.
Arrabales, R. and Sanchis, A. 2008, "Applying machine consciousness models in autonomous
situated agents", Pattern Recognition Letters, vol. 29, no. 8, pp. 1033-1038.
Arrabales, R., Ledezma, A., and Sanchis, A. 2007, "Modeling Consciousness for Autonomous
Robot Exploration", Lecture Notes in Computer Science, vol. 4527, pp. 51-60.
Arrabales, R. and Sanchis, A. 2006, "A Machine Consciousness Approach to Autonomous
Mobile Robotics", 5th International Cognitive Robotics Workshop. AAAI-06.
Atkinson, A.P., Thomas, M.S.C., and Cleeremans, A. 2000, "Consciousness: Mapping the
Theoretical Landscape" in Trends in Cognitive Sciences. vol. 4, no. 10, pp. 372-382.
Baars, B.J. 2002, "The conscious access hypothesis: Origins and recent evidence", Trends in
Cognitive Science, no. 6, pp. 47-52.
Baars, B.J. 1997, "In the Theatre of Consciousness: Global Workspace Theory, A Rigorous
Scientific Theory of Consciousness", Journal of Consciousness Studies, no. 4, pp. 292-309.
Dennett, D.C. 1991, Consciousness Explained, Little, Brown and Co, Boston.
ARRABALES ET AL
20
Goertzel, B. 2008. "Achieving Advanced Machine Consciousness though Integrative, Virtually
Embodied Artificial General Intelligence", Proceedings of the Nokia Workshop on Machine
Consciousness. pp. 19-21.
Haikonen, P.O.A. 2007, Robot Brains. Circuits and Systems for Conscious Machines, John Wiley
& Sons, UK.
Haikonen, P.O.A. 2008, "The Role of Qualia in Machine Consciousness; Nothing or Everything?
(The Answer is YES). Proceedings of the Nokia Workshop on Machine Consciousness, pp.
44-45.
Hofstadter, D. 1995, "The Copycat Project: A model of mental fluidity and analogy-making" in
Fluid Concepts and Creative Analogies Basic Books.
Holland, O. 2007, "A Strongly Embodied Approach to Machine Consciousness", Journal of
Consciousness Studies, vol. 14, pp. 97-110.
Microsoft Corp. 2006, Microsoft Robotics Studio. http://msdn.microsoft.com/robotics/.
Minsky, M. 1988, Society of Mind, Simon & Schuster, New York.
Nielsen, H.F. and Chrysanthakopoulos, G. 2006, Decentralized Software Services Protocol -
DSSP, Microsoft Corporation.
Nii, H.P. 1986, "Blackboard Application Systems, Blackboard Systems and a Knowledge
Engineering Perspective", AI Magazine, vol. 7, no. 3, pp. 82-107.
Ramamurthy, U., Baars, B., D'Mello, S.K., and Franklin, S. 2006, "LIDA: A Working Model of
Cognition", Cognitive Modeling, eds. F.D.M. Danilo Fum & Andrea Stocco, Edizioni
Goliardiche, pp. 244.
Richter, J. 2006, "Concurrent Affairs: Concurrency and Coordination Runtime", MSDN
Magazine, no. 10.
Shanaham, M. 2005, "Consciousness, Emotion, and Imagination. A Brain-Inspired Architecture
for Cognitive Robotics", AISB Workshop Next Generation Approaches to Machine
Consciousness.
Shanahan, M. 2006, "A cognitive architecture that combines internal simulation with a global
workspace", Consciousness and Cognition, vol. 15, no. 2, pp. 433-449.
Shanon, B. 2008, "A Psychological Theory of Consciousness", Journal of Consciousness Studies,
vol. 15, pp. 5-47.
Singh, M.P. & Huhns, M.N. 2005, Service-Oriented Computing. Semantics, Processes, Agents.
John Wiley and Sons.